Machine learning in trading: theory, models, practice and algo-trading - page 2061

 
Rorschach: I want to "scientifically" prove the significance of the result. For this purpose I use the law of large numbers and 4*sko. Can this be used for TS with unequal SL and TP, and how many trades are needed?

Simulation for the first system from the archive. Red lines worst and best results for 100,000 iterations. The lower bound goes to the plus after 2500 trades (40 years).

I calculate by formula sqrt(1/572*(514*(29.95-6.64)^2+58*(-200-6.64)^2)) = 69.4

3796.3/69.4=54.7, it seems to be a significant result

or do I have to divide the expected payoff by the sko?

I added sko to the model:

10000-3754=6246 //max deviation for 572 trades over 100000 iterations

69.4*sqrt(572)=1659.8 //sco for 572 trades

6246/1659.8=3.76

I think I've got the wrong calculation.


The simulation for expectation=0. In general, the result is within the error.


 
mytarmailS:

The solution is this: we make an extremum pattern and look at the correlation between the current price and the pattern...

How many bars on both sides of the pattern? Have you calculated what percentage of movement remains after identifying the template until the next extremum ZZ?

 
Evgeniy Chumakov:


For a zigzag, it would be enough to teach the neural network to predict the length of the shoulder - longer than the previous one or shorter (as the target needs long ones).

My system works like that, the prediction is given at the moment when a new ZZ segment appears. Like all systems - the financial result depends very much on market conditions, i.e. whether there will be many ZZ's, exceeding a given threshold. If the threshold is small, there will be many profitable trades, but there is a problem with stops - I put them at a point where a new interval ZZ appears, so far I have not found any other variant for the initial entry.

 
Aleksey Vyazmikin:

How many bars on either side of the pattern? Calculated what percentage of movement remains after the pattern is revealed to the next extremum ZZ?

10 on each side (I took it from BALDA)

No I haven't, you can count it by the way, it's interesting to see...

I'm still stuck on my patterns, it's an interesting thing

 
mytarmailS:

10 on each side (I took it from BALDA)

No I didn't, you can count by the way, it's interesting to see...

I'm still on my templates, interesting thing turned out

If I manage to determine the depth of correction as a percentage of the previous interval of ZZ, it may be interesting. In the meantime, I should consider an exit point at the MA (96-128) or when a new interval of ZZ appears.

While I'm busy straightening the strategy on the Mach - I'm doing it for the article.

 
Aleksey Vyazmikin:

If we manage to determine the depth of the correction as a percentage of the last interval of ZZ, it might be interesting. In the meantime, we should consider the exit point by MA (96-128) or at the appearance of a new interval of ZZ.

While I am busy with straightening the strategy on my Mach, I am doing it for my article.

I do not use ZZ at all

 
mytarmailS:

I don't use zz at all

It is clear that you identify and small movements, but it is essentially the same ZZ with a smaller value.

 
mytarmailS:

Trained.

Gray is the original data, blue is from the model

I also inserted a piece of trail to show how the model fades out on the new data


Do something already. I want a lamba, I can't get enough of it. There's no money for it in the factory 😔
 
usually strategies are tested by monte carla, if you need it badly
 
mytarmailS:

Oh fuck.... If you knew what's in my vulnerable soul )))) I don't believe in anything ...

This hat understands the flat, but it's sad about the trend.


Try to filter by time, for example in Asian / Pacific session, it works in the flat