Machine learning in trading: theory, models, practice and algo-trading - page 236

 
SanSanych Fomenko:

What is a U-turn? One bar as in ZZ?

Yes, one bar.
 
mytarmailS:
Yes, one bar.

How about the way I wrote above?

We take the ZZ and set a parameter in it, for example, 50 pips. Let's build. On the real chart, 50 pip reversals are very rare when the parameter is 50 pips. Before the reversal and after the reversal to the 50 pip boundary, we mark it as the previous reversal...

If we build a PZ with the parameter of 70 pips and set the boundary at 50 pips, there will be even more such marked bars as "reversal" than in the previous case.

PS.

I've been cooperating on this topic. I failed to build a decent model. But this failure cannot be considered proof that the idea itself is wrong - just a specific case.

 

Trained RF on cumulative random,

This random looks like this :

й

anyone distinguish from the price ???

target reversals

a reversal downwards is a point which is more than 10 previous points and more than 20 subsequent points

Let's check the model trained on the randoms on real quotes

One and the same model has been trained several times (trials) and tested in trading.

RTS instrument

i took median price as a price (high+low)/2

stop and take was the same - 50 ticks

the commission is taken into account, so actually the stop is even less than the stop

Stops and tees are fixed, nothing moved, just a stupid pattern, stupid TS

trial one

ф

trial two

я

trial three

ц

trial four.

у

and then it goes on like this....

training code
Files:
zzz.txt  1 kb
 

mytarmailS:
Yes, one bar.

========================================

1-2 bars before the top of the ZZ and 1-2 bars after. We need to use other models.

Deep NN training consists of two steps, you know.

1.Pre-training without a teacher on as large a set of unlabeled data as possible (I take 5-6 thousand bars) and

2. Tweaking on a smaller sample only near the tops of (1-2)bars.

The result is better than when you race the whole sample.

Good luck

 
mytarmailS:

Trained RF on cumulative random,

looks like this :

anyone distinguish from the price ???

target reversals

A downwards reversal is a point which is greater than 10 previous points and greater than 20 subsequent points

Let's check the model trained on the randoms on real quotes

One and the same model has been trained several times (trials) and tested in trading.

RTS instrument

i took median price as a price (high+low)/2

stop and take was the same - 50 ticks

the commission is taken into account, so actually the stop is even less than the stop

Stops and tees are fixed, nothing moved, just a stupid pattern, stupid TS

trial one

trial twoc

trial three

assay 4

and more like this....

training code
So what's not on the real quotes?
 
Vladimir Perervenko:
So, what's not on the real quotes?
Just making pictures now
 

now how exactly the same model(those with the same parameters) recognized real quotes and also trained on real quotes, not on random

1

й

2

ф

3

я

4

у

Let me explain one more time

In the first case the model was trained on the random and tested on real quotes, the pictures show trading on real quotes

In the second case the model was tested and trained using real data

 
Vizard_:
All you know, why did not ask is not clear. Just at what they ask and train, and no alchemy. Metric, of course to the amateur, I usually do not yuse her.

My process looks like this. And to train one or a bunch of models is everyone's choice))). With this task in this branch I think a lot of good to cope ...


I asked what kind of model, perseptron, forest, etc.. But the essence you said, trained with crossentropy
 
mytarmailS:

now how exactly the same model(those with the same parameters) recognized real quotes and trained on real quotes too and not on random

No one even commented?

Personally I'm shocked to say the least...

 
mytarmailS:

No one even commented on it?

Personally, I'm at the very least freaked out...

Well, what? all bad, apparently? to beat the market in the form of buy and hold has failed.